In the financial industry, especially in the securities field, the application of AI large models is also gradually being explored. This article will delve into the application prospects of AI large models in brokerage user operations, from customer acquisition, activation, retention, revenue to recommendation, and analyze how to optimize user operation strategies and improve user experience and business efficiency through AI technology.
AI large models are developing rapidly, the scale of parameters is jumping rapidly, the cost of inference training is decreasing, and AI large models have begun to be widely used in the fields of medical care, education, office, and media creativity, and people are beginning to believe that it will become an important infrastructure like electricity and networks in the future.
In the financial industry,Banks with strategic vision and capital strength have taken the lead in competing for AI talents——From large state-owned banks such as ICBC and CCB, to joint-stock banks such as China Merchants Bank, and even city commercial banks such as Bank of Chengdu, they have all launched systematic AI talent recruitment plans this year. From the perspective of the current recruitment direction, the bank’s AI talent layout is mainly in robo-advisory, risk assessment, customer service and other aspects.
Figure 1: AI talent recruitment plan
Compared with banks, the scenarios of AI application in securities companies are different and different.The scenario of securities companies is mainly in the generation of research reports, robo-advisory and risk control compliance.Among them, research report writing is highly dependent on text processing capabilities, which is most in line with the technical characteristics of large language models, and has become the application direction with the highest penetration rate.
However, in the field of user operation of securities companies, the application of AI is still in the exploration stage, and most institutions are still in the traditional manual or digital marketing paradigm. However, the user operation of “attracting, activating, and converting” has always been crucial for securities companies. How to use AI capabilities to reconstruct the customer journey, improve operational efficiency, and achieve a leapfrog improvement in user operation capabilities?
Let’s start with the question: the current situation of brokerage user operations under the AARRR model
The AARRR model was first proposed by Silicon Valley venture capitalist Dave McClure in 2007 and has since been widely used in the Internet industry as a commonly used operating model in “user operations”. The model divides each step of user operation from five links: customer acquisition, activation, retention, revenue, and recommendation.
Figure 2: AARRR funnel
1. Acquisition
For securities companies, the goal of the customer acquisition stage is to open an account for customers. At present, the customer acquisition of brokerage business mainly relies on four mainstream models: bank-securities cooperation, telemarketing, account manager network development and Internet advertising.
Various models show differentiated characteristics in practice:Bank-securities cooperationThe channel has the advantage of a large customer base, but in fact, after the customer opens an account,The proportion of customers participating in investment transactions is low；TelemarketingThe model takes the activity of customer expansion or channel resources as the main source of leads, and its effectiveness is subject to the effect of telephone lead quality and outbound conversion rate, which is obviousManpower-intensiveFeatures;Account ManagerThe network model has advantages in quality, but its business growth is highly dependent on the personal resource endowment of employeesThe cost of manpower superposition is high and large-scale replication is difficultbottleneck; Internet advertising follows the logic of traffic purchase, and has had large dividends in the Internet population growth stage and the early stage of industry competition by seizing search rankings and cooperating with leading investment platforms.However, with the continuous rise of traffic costs and the problem of too homogeneous marketing content, this model is also facing the pressure of high costs and inverted revenues.
2. Activation
The target of activation is the customer’s first key action, which refers to the customer’s direct value action, including deposits, stock trading, fund purchases, etc.
The brokerage is still inThe account manager assumes the primary activation responsibilities, through one-on-one communication, guide new customers to complete the first operation. Although this model can establish interpersonal trust, it is limited by the service radius and can often only cover the customer groups independently developed by account managers, resulting in a large number of customers who are not self-developed relationships by account managers lack systematic activation guidance. On the other hand, in terms of activation methods, customers from different channels should have certain differences in activating their products, such as bank-securities cooperative customers tend to make their first deposit through fixed income products, while Internet channel customers are more likely to be activated by intelligent stock selection tools. But at the executive level,Brokerages have not yet established a standardized matching mechanism for channel characteristics and activation strategies, and still rely on account managers for personalized adaptation based on experience.Due to the differences in the marketing capabilities of various agencies and personnel, the experience-driven activation effectIt is difficult to achieve balanced development.
3. Retention
The goal of retention is the retention of clients’ funds at the brokerage and the retention of trading clients.
What does a product manager need to do?
In the process of a product from scratch, it is not easy to do a good job in the role of product manager, in addition to the well-known writing requirements, writing requirements, writing requirements, there are many things to do. The product manager is not what you think, but will only ask you for trouble, make a request:
View details >
On the one hand, customers independently adjust their position structure based on factors such as personal investment motivation (such as long-term asset allocation needs), market fluctuations (bull market capital precipitation effect), and hot information stimulation (theme investment opportunities) to form natural retention. On the other hand, there is active marketing by employees, where account managers guide customer funds to stay in the company’s account by recommending specific financial products or trading strategies.
thisIt is more applicable when the market is good, but the loss is higher when the market is down。 At the same time, because the demand for personalized services for customer risk appetite and trading habits relies on manual experience, the granularity of the service is relatively extensive. Too much reliance on “people” services will also be limited by human efficiency and numbers, which will restrict the overall scale expansion of the company.
4. Revenue
Securities brokerage has long relied heavily on channel income such as trading commissions and margin spreads, and such business models have occupied a central position over the past few decades. Although the industry has generally recognized the inevitability of the transition to wealth management – especially for demand-oriented businesses such as buy-side investment advisory. However, at the practical level, the strategic investment in service capacity building is still lagging behindprofessional services represented by investment consulting services and asset allocation research started late and had a long path of personnel training; the resource investment for the buyer investment advisory model is relatively insufficient; The speed of building an intelligent service system has not yet matched customers’ expectations for personalized and real-time investment advice.
Especially in the process of business promotion and service, value transmission is highly dependent on the subjective initiative of front-line employees, when account managers are limited by service capabilities (such as energy allocation, professionalism) and cannot take the initiative to recommend,A large number of medium and long-tail customers are actually in the service blind spot, neither understanding the diversified product matrix that brokerages can provide, nor perceiving the decision-making support value brought by intelligent tools。 This separation of “service supply” and “customer cognition” has also become a constraint on the upgrading of business revenue structure.
5. Referral
Referral refers to the introduction of new customers by old customers. The offline end mainly relies on account managers to carry out referrals through interpersonal communication, while online channels generally lack systematic fission operations. Although a few brokerages with resource synergy advantages, such as insurance brokerages, have laid out social fission marketing earlier, they have achieved rapid growth in customer scaleHowever, after the introduction of regulatory requirements that “marketing gifts must be strongly related to the securities business”, the effectiveness of traditional fission incentive methods (such as physical gifts and universal points) has been significantly weakened.
At present, the industry is actively exploring innovation paths under the compliance framework, such as designing securities-related rights and interests such as investment education courses and smart tool experience vouchers as fission incentive targets to balance regulatory requirements with user attractiveness.
Several types of application directions of AI large models
AI large models refer to models with massive parameters and complex architectures for deep learning tasks, which have gone through development stages such as single-language pre-trained models, multi-language pre-trained models, and multi-modal pre-trained models. It is usually based on deep learning and neural network technology, which can handle more complex tasks and has stronger generalization capabilities. According to the application direction, there are mainly the following categories:
1. Natural Language Processing (NLP)
The natural language processing (NLP) large model isEnable machines to perform tasks such as text understanding, information extraction, automatic translation, and sentiment analysis, enabling natural human-computer interaction.
It is mainly used in dialogue, text generation, translation, information extraction and summarization. The problem with NLP large models is that in the absence of real-world basis, they may generate false or misleading content, that is, “hallucinations”. At the same time, reasoning ability relies on data and does not have common sense like humans.
2. Multimodal (text + image + video + audio)
A multimodal large model is an AI model that can process and understand different types of data (such as text, images, videos, audio, etc.) at the same time.It can perform cross-modal tasks, such as graphic understanding, video generation, speech recognition synthesis, etc。
It is characterized by “leapfrogging”, which can be freely created through text, transcending the general single modality of only reading text or listening to audio. When processing speech information, even relatively complex situations such as accents and noises can be recognized; When processing video information, it has also changed from analyzing a single frame of video to understanding the entire video. The main problems are unstable generation results, high computational costs, and the accuracy of image understanding needs to be improved.
3. Search and code generation
Search and code large models refer to large-scale artificial intelligence models specifically used for information retrieval and code generation. They are based on large-scale pre-trained language models, combined with search engines, code corpora, and specific optimization techniques to achieve efficient information query, accurate code completion, and automated programming support.
The search for AI large models canTell the answer directly after searching for a question or keyword, rather than just providing pages related to the search term for users to find the answer themselves, as in the past.
Code generation large models are large models that make programming more efficient and can helpComplete code, fix bugs, and optimize performance。 It can improve development efficiency and reduce repetitive labor, but it still relies on manual inspection and judgment.
4. Strengthen learning and decision-making optimization
Reinforcement learning and decision optimization large models are AI models specifically designed for automated decision-making, strategy optimization, and long-term planning. They are widely used in scenarios such as robot control, autonomous driving, intelligent trading, and industrial scheduling, enabling AI to have human-like decision-making capabilities.
Reinforcement learning is learning that allows AI to practice “trial and error”.By allowing AI to reward or punish, let AI explore the optimal strategy by itself。 Reinforcement learning and decision optimization large models can be used for automated decision-making, strategy optimization, quantitative trading, robot control, etc., and it is good at finding optimal solutions in environments without standard answers, and can even surpass human intuition and experience. Reinforcement learning and decision optimization require extremely high training data, high computational cost, long training time, and strong dependence on the training environment.
5. Knowledge graph model
The knowledge graph model combines knowledge graph with large-scale pre-trained language models to enhance AI logical reasoning, knowledge storage, and question-answering capabilities. It plays a key role in tasks such as information retrieval, intelligent question answering, recommendation systems, and automated reasoning.
The knowledge graph isA large AI model that stores and represents knowledge in a structured way, which builds a knowledge network through “entities” and “relationships”, allowing machines to understand and reason information like humans.For example, A is B’s mother, B is C’s father, and the knowledge graph can deduce that A is C’s grandmother. Knowledge graphs can be multi-modal fusion and structured data storage, which is suitable for data or professional knowledge search and answering, but there are certain limitations in terms of construction cost and generalization of knowledge.
How to use AI large models to solve the problem of brokerage user operations
In order to realize the application of AI large models in securities brokerage business, the overall framework design can be divided into the following four levels, progressing layer by layer to ensure that the technology implementation is accurately matched with business goals.
Figure 3: Architecture design
The data layer is the foundation,Provide all the data needed to support AI large models, applications, and business goals。 It includes four types of data: users, content, behavior, and transactions, which can be summarized as who, what they watched, what they did, and what they bought. Who is the user (static information of the user, such as identity information), what is seen (information about the user’s choice and consumption, such as information, research reports, videos, stock analysis, etc.), what is done (the user’s operation and interaction process, such as what is clicked, what is browsed, how long it stays), what is bought (actual buying and selling decisions, and data calculated from buying and selling decisions, such as transaction frequency, etc.).
The AI large model layer processes data and generates intelligent decisions, and it is on top of the data layer.By calling the data of the data layerAfter unified processing, cleaning, and integration, the data of a certain type of feature is formed into a unified dataset and then trained and optimized, and patterns and trends in the data are found, which are then used for prediction and decision-making at the application layer.
The application layer is the implementation of AI large model results in business scenariosIt does not perform calculations, but issues instructions and retrieves results from the AI large model layerThe AI large model layer then extracts the underlying database through the instructions of the application layer and calculates, and outputs the results to the application layer after obtaining the results of the application layer instructions.The application layer uses the instruction results to reach customers, and finally uses the data results to feed back to the data layer, providing more data for the AI large model layer for later model optimization, forming an automatic improvement cycle.
As a strategic layer, the business target layer sets goals for the application layer, and the actions of the application layer need to be carried out around the goal, affecting the operation and optimization direction of the AI large model, so that the AI can closely focus on the business goal during training.At the same time, adjust business goals in a timely manner according to the feedback from the AI layer and the application layer, such as the AI layer finds that the account opening effect brought by the live broadcast is better than the article, and the investment in live broadcast should be increased in the business goal to ensure that the business direction is correct.
The business target layer, application layer, AI large model layer, and data layer form a top-down and bottom-up dual cycle, and continuously optimize to form a closed loop.
Based on the theoretical framework of the AARRR model and relying on the technical support of five types of AI large models, this study will further propose an optimization path around the current situation and problems of the whole process of user operation.
1. Customer acquisition stage
As the first link of the securities brokerage business, customer acquisition can achieve breakthroughs in three dimensions: customer acquisition efficiency, material delivery and cost control through the application of AI large models:
(1) Optimization of customer acquisition efficiency
In the process of customer acquisition, account managers will obtain a large number of customer leads, in order to make the leads have higher conversion efficiency,First, AI agents can be built based on knowledge graph technology, integrating multi-source heterogeneous data (including customer lead source channels, clicked material information, etc.) to build accurate potential customer portraits and analyze the customer’s possible account opening interest points.For example, customers who click in through the new customer gift package advertisement focus on bringing customers the feeling of giving benefits during marketing.
and then passedAI agents replace account managers or call centers to complete basic customer contact tasks, combined with reinforcement learning and decision-making optimization algorithms, continuously optimize communication strategies to improve customer conversion rates.
(2) Optimization of customer acquisition materials
The materials for customer acquisition are generally written and recorded by employees themselves, combined with AI large model technologyRelying on NLP (natural language processing) and multimodal large model technology, it generates professional and attractive graphics and videosand other customer acquisition materials to meet diversified marketing needs. For example, digital human technology, which is now widely used in the field of education, only requires a person’s image and AI large model to write scripts, which can generate multiple videos in one day.
After the AI large model generates content, brokerages can build a matrix operation strategy by establishing multiple accounts on social short video platforms, breaking through the limitations of platform algorithms in the form of “high-frequency updates” and “large-scale coverage”, maximizing platform traffic, and enhancing the brand awareness and customer reach effect of brokerages.
(3) Customer acquisition cost control
When delivering, NLP large models can analyze keyword preferences based on user portraits and behavior data, optimize advertising copy and creativity, and improve click-through rates and conversion effects.Through AI algorithms, real-time analysis of delivery data can dynamically adjust bidding strategies and target group selection, maximize the ROI of advertising and effectively control the cost of customer acquisition.
Through the above strategies, the application of AI large models can not only improve the efficiency of customer acquisition and material quality in the securities business, but also significantly reduce the cost of customer acquisition, and provide sustainable technology drivers for business growth.
2. Activation phase
The core goal of the activation stage is to guide customers to take their first key actions and enhance their continuous attention to securities products and services.
(1) Service radius expansion
The brokerage business customers of securities firms are generally divided into customers who establish a linked relationship independently developed by employees and customers who open accounts independently and do not have a linked relationship with employees.For customers independently developed by employees, the customer relationship network can be analyzed through knowledge graph technology, providing accurate personalized service materials for salespeople and improving customer interaction efficiency and satisfaction. For example, for high-net-worth clients developed by employees, the knowledge graph and NPL can introduce the company’s financial products and specific services suitable for qualified investors.For customers who do not have employee connections, use NLP technology to analyze user interests and preferences, recommend generic or personalized content to guide them to complete key actions.
(2) Intelligent service strategy
For customers from different channels, AI technology analyzes user behavior data (such as account opening channels, browsing history after entering the APP, etc.), and recommends information and research reports that match their interests.NLP technology can also be used to generate a universal account opening “new friend welcome guide”, as well as differentiated customers with different channels and different portraits to use different activation tools and products, such as fixed income and currency products for activation for bank channel customers or middle-aged and elderly customers.
Through the above strategies, the application of AI large models can effectively improve users’ experience in the activation stage, enhance their interest and dependence on securities products and services, and lay a solid foundation for subsequent retention and conversion.
3. Retention phase
The key to improving user retention is to continuously monitor user behavior through AI technology and implement precise intervention strategies to achieve the goals of capital retention and transaction customer retention. Specifically, it can be carried out from three aspects: passive strategy, active strategy and churn warning:
(1) Passive strategy
Passive strategy refers to the client’s passive acceptance of the broker’s services.Brokerages use NLP large models to identify the financial content browsed by users, through word segmentation, sentiment analysis and other technologies to analyze the financial information, research reports, investment advisory analysis and other content browsed by users, extract keywords,Combined with the knowledge graph model of securities firms, the user’s behavior is related to the larger financial concept systemFor example, users who browse CATL, the system marks them as “new energy interest”, and in the future, when new energy-related financial reports and industry trends are released, AI will give more recommendation weight to this customer. Combined with the AI model of reinforcement learning, the recommendation strategy is continuously adjusted to reduce content recommendations with low click-through rates and short reading time, thereby improving user experience and customer retention.
(2) Active strategy
If the user actively needs brokerage services, they canCombined with the construction of an intelligent search system based on the knowledge graphto improve the accuracy and efficiency of financial content retrieval.It can also build a Q&A agent robot, which can allow customers to interact and answer questions independently, which can also enable employees to quickly query and answer customer questions. NLP technology supports natural language queries to help users quickly obtain the information they need, improving customer interaction and satisfaction.
(3) Loss early warning and intervention
In order to prevent the loss of customer funds, machine learning, reinforcement learning, and decision optimization large models can be combined to predict the time point when users may churn, and interventions can be formulated in combination with reinforcement learning. The machine learning model is used to find key features, calculate the probability of key features and user churn, and then intervene and optimize through reinforcement learning and decision optimization.For example, when AI calculates that the market has fallen by 10%, 70% of customers in the past have chosen to sell and no longer buy, and the churn rate has become higher.
Through the above strategies, the application of AI large models can effectively improve user retention, reduce the risk of churn, and enhance users’ long-term dependence and trust in brokerages by providing users with the products and services they need.
4. Income stage
The key problem that needs to be solved in the revenue stage is to transform into buy-side investment advisory, from “product-oriented” marketing to “customer demand-oriented”, improve the overall quality of the company’s investment advisory services, increase the reach and conversion rate of financial products, enhance customer cognition, and encourage customers to actively choose products that suit them.
(1) Improve investment advisory capabilities and work efficiency
NLP and knowledge graph large models can help investment advisors handle investment research work, and at the same time, combined with reinforcement learning and decision optimization, they can generate intelligent investment recommendations for customers.Suppose an employee with investment advisory qualifications and capabilities needs to serve 500 customers, combined with the AI large model, it can automatically generate information interpretation and position adjustment suggestions for customers more quickly, while employees only need to serve the most important customers in a refined manner.
(2) Understand customer needs
In order to allow customers to better understand the product, find the financial products they need. Search large models, reinforcement learning, and decision optimization can be used to understand users’ financial needs. For example, when a customer searches for “the impact of interest rates on funds”, the system can judge whether customers are interested in bond funds, and then when generating materials to reach customers, AI can find out what forms are easier to impress customers. In this way, it not only promotes the products that users are interested in, but also tells the products clearly. It not only improves the professionalism of investment advisory, but also lowers the threshold for customer understanding, so that the transformation of buy-side investment advisory can be truly implemented, and the upgrade of the revenue stage is promoted.
5. Referral stage
In the recommendation stage, brokerages should promote the willingness and efficiency of old customers to invite new customers, while meeting compliance requirements.
(1) Design activities and identify key users
When designing activities, passMultimodal models generate engaging campaign materials(such as graphics, texts, videos, etc.) to enhance the reach effect of activities. Based on knowledge graph and NLP technology, analyze users’ social relationships and influence, and identify high-value key customers (such as active users, opinion leaders, etc.). Develop exclusive referral incentive strategies for key customers to maximize their communication effectiveness and conversion rate.
(2) Strengthen learning and decision-making optimization, and continuously optimize the recommended population
If the AI large model finds that “friends of high-net-worth customers are more likely to open accounts” and “young customers are more willing to share but have a low conversion rate”, then the strategy can be adjusted in the future.For high-net-worth customers, referral programs related to their high-end investment rights and interests are introduced, and young customers use a more lightweight social fission method.
To verify which type of reward is more engaging, use reinforcement learning combined with A/B testing to continuously test adjustments. In terms of rewards for attracting conversions: if AI finds that customer A likes ETF investment, then equity is more inclined to “ETF investment courses and commission discounts”, customer B pays attention to short-term trading, and AI can push “short-term trading strategy” conversions; On the path of transformation, users can still analyze the user’s fission behavior through reinforcement learning combined with A/B test to find out whether it is better to share in the live broadcast room or in the circle of friends. Through the above strategies, the AI large model improves the effect of the recommendation stage, realizes user fission growth and the expansion of brand influence.
6. Compliance guarantees
Compliance is the core prerequisite for the development of securities business and the bottom line of business development. The application of AI large model technology is significantly innovative, so compliance management needs to be paid great attention to in business development. Based on the AARRR model, the core process can be summarized into key links such as data acquisition, content production, push marketing, and transaction conversion, and each link needs to establish a complete compliance risk control mechanism to ensure that business development meets regulatory requirements.
(1) Compliance with data acquisition
Data acquisition needs to comply with the Personal Information Protection Law to ensure that the collection, storage, and processing of customer data are legal, and before AI uses customer data, it is necessary to obtain explicit authorization from customers, and some sensitive data should also be encrypted and stored to restrict AI’s access rights.
(2) Compliance with content production
All content generated by AI large models (such as investment research reports, market analysis, etc.) must be reviewed by professional compliance personnel to ensure that the content complies with laws, regulations and regulatory requirements, and avoid the dissemination of content that does not meet compliance requirements.
(3) Compliance of push marketing
In the process of content push, it is necessary to ensure that the recommended content matches the user’s risk tolerance and other customer risk appropriateness to avoid compliance risks caused by improper recommendation. In view of the possible hallucination problems of AI large models, it is more necessary to clearly inform customers of the basis for pushing (such as AI judgment based on user behavior data, market trends, etc.), and provide customers with independent decision-making space to avoid the problem of algorithm black box and protect users’ right to know and choose.
(4) Compliance of transaction conversion
AI prediction and recommendation strategies may cause users to overtrade, and in terms of risk and compliance management, it is necessary to set reasonable transaction frequency thresholds to avoid user losses or market fluctuations caused by excessive trading. Avoid AI using information asymmetry or algorithmic advantages to manipulate the market, and maintain market order and user rights.
(5) Compliance with algorithm management
The algorithm of the AI large model runs through the entire operation process, and its compliance management needs to strictly follow the national “Regulations on the Administration of Internet Information Service Algorithm Recommendation” to ensure that the algorithm design is fair, ethical, transparent and explainable, uphold the concept of upward goodness, and conform to the mainstream value orientation.
Through the above compliance safeguards, the application of AI large model technology can improve business efficiency while ensuring business compliance, providing customers with safe and transparent services.
epilogue
AI has strong application potential in the operation of brokerage users, but there will be many challenges in the actual implementation process: in terms of the combination of technical scenarios, the two-way cycle of “AI large models should empower the business, and the business should feed back the AI large model” to ensure that technology and data can fully support the realization of business goals; Secondly, in terms of the division of human-machine responsibilities, AI large models can replace humans to do massive data and complex calculations and make automated and intelligent decisions, but in the field of securities with high compliance requirements, people still need to control professional judgment, compliance risk control, and user trust. Third, in terms of compliance business, it is necessary to fully comply with policies and regulatory requirements, protect the interests of investors, avoid inducing customers, ensure that information is true, accurate and complete, avoid misleading impact on investors, and effectively safeguard market order and customer rights and interests.
With the gradual improvement of AI large model technology, the wealth management business of brokerages will achieve intelligence, personalization and automation, promote the development of wealth management of securities companies from “manual experience-driven” to “data intelligence-driven”, and promote the marketing model from “product-oriented” to “customer demand-oriented”.
Facing the future, do a good job in the present.
